Latest / The AI Practitioner Brief / Meta's Free Model, Google's TPUs, and CoreWeave's Debt — October 16, 2026

Transcript

Meta's Free Model, Google's TPUs, and CoreWeave's Debt — October 16, 2026 · The AI Practitioner Brief

• 0:00
  [music]
• 0:05
  SPEAKER_00: Welcome to The AI Practitioner Brief,
• 0:08
  SPEAKER_00: a podcast from AIPractitioner.substack.com,
• 0:11
  SPEAKER_00: where I turn the latest AI and tech news into practical insights
• 0:15
  SPEAKER_00: for AI practitioners.
• 0:17
  SPEAKER_00: I'm Lina Faik,
• 0:18
  SPEAKER_00: an ML engineer,
• 0:19
  SPEAKER_00: consultant, and lecturer in advanced machine learning at Sorbonne University.
• 0:23
  SPEAKER_00: It's Friday, October 16th.
• 0:25
  SPEAKER_01: Friday already.
• 0:26
  SPEAKER_01: Let's get into it.
• 0:27
  SPEAKER_00: With me is Jimmy, my co-host on the brief who asks the questions I'd skip past.
• 0:32
  SPEAKER_00: Today has a real contradiction at its center.
• 0:34
  SPEAKER_00: Meta just gave away one of the most capable AI models ever built.
• 0:38
  SPEAKER_00: Completely free.
• 0:39
  SPEAKER_00: That happened the same day Wall Street started worrying about a company's AI debt.
• 0:44
  SPEAKER_00: That company rents out the chips that power models exactly like this one.
• 0:48
  SPEAKER_00: Investors are now pricing in the chance it can't pay that debt on time.
• 0:51
  SPEAKER_01: Free model shaky chip bill, where do we start?
• 0:54
  SPEAKER_00: We'll start with the headlines, then three bigger stories.
• 0:58
  SPEAKER_00: Meta's new open model is first,
• 1:00
  SPEAKER_00: then we'll get into why five top AI models cost almost four times a part to run,
• 1:04
  SPEAKER_00: despite scoring about the same.
• 1:07
  SPEAKER_00: And we'll end on why the company underneath a big chunk of
• 1:10
  SPEAKER_00: the AI boom's compute is suddenly scaring its own bond investors.
• 1:13
  SPEAKER_01: Start with the AI stuff.
• 1:15
  SPEAKER_00: Liquid AI,
• 1:16
  SPEAKER_00: a startup that builds small,
• 1:18
  SPEAKER_00: efficient models, released a decision model called D1 Ultra today.
• 1:21
  SPEAKER_01: What does a decision model even do?
• 1:24
  SPEAKER_00: Instead of writing out an answer,
• 1:25
  SPEAKER_00: it reads a question against a photo or document and just returns a probability,
• 1:30
  SPEAKER_00: no generated text at all.
• 1:31
  SPEAKER_00: Think spotting a defect on a circuit board or checking a damaged claim against
• 1:35
  SPEAKER_00: a customer's photo.
• 1:36
  SPEAKER_01: Any number behind best, or is that just marketing?
• 1:39
  SPEAKER_00: Independent testing put it at 65 on the Artificial Analysis Decision Index.
• 1:44
  SPEAKER_00: That's 8 points past the previous leader.
• 1:47
  SPEAKER_00: It's priced at under 5 cents per million input tokens, with nothing charged for output.
• 1:52
  SPEAKER_01: Cheap and apparently checked.
• 1:53
  SPEAKER_01: What else?
• 1:54
  SPEAKER_00: A new analysis found that 1% of companies account for 80% of all spending on OpenAI
• 2:00
  SPEAKER_00: and Anthropic.
• 2:00
  SPEAKER_01: That's a thin customer base to build a business on.
• 2:04
  SPEAKER_00: Even as usage keeps rising,
• 2:05
  SPEAKER_00: spending per employee fell a little over 3% because prices are dropping fast.
• 2:10
  SPEAKER_01: So usage is up, but the bill is shrinking.
• 2:12
  SPEAKER_00: That's the AI side.
• 2:14
  SPEAKER_00: Elsewhere in tech today,
• 2:15
  SPEAKER_00: Tesla unveiled its second Roadster this week,
• 2:18
  SPEAKER_00: launching it up a ramp into a controlled landing using rocket thrusters borrowed
• 2:22
  SPEAKER_00: from SpaceX.
• 2:23
  SPEAKER_01: Rocket thrusters.
• 2:24
  SPEAKER_01: On a car.
• 2:25
  SPEAKER_00: The fastest version reportedly hits 60 miles an hour in well under a second.
• 2:30
  SPEAKER_00: Musk put the top speed past 230 miles an hour.
• 2:33
  SPEAKER_00: Production isn't due until 2028, and Tesla still won't say what it costs.
• 2:38
  SPEAKER_01: Convenient timing on that last part.
• 2:40
  SPEAKER_00: Separately, tech companies are increasingly building their own internal vibe coding apps.
• 2:45
  SPEAKER_00: Vibe coding just means describing what you want in plain English and letting
• 2:48
  SPEAKER_00: an AI write the code.
• 2:50
  SPEAKER_01: Instead of buying one off the shelf?
• 2:52
  SPEAKER_00: Right, and it's getting higher adoption than vendor tools,
• 2:54
  SPEAKER_00: partly because the company's own data never leaves the building.
• 2:58
  SPEAKER_01: Makes sense for anything sensitive.
• 2:59
  SPEAKER_00: One more.
• 3:00
  SPEAKER_00: Rippling,
• 3:01
  SPEAKER_00: which sells HR and payroll software,
• 3:03
  SPEAKER_00: ditched automated outbound email for dedicated human sales reps.
• 3:06
  SPEAKER_00: Conversion jumped from around 1% to between 3% and 7%.
• 3:10
  SPEAKER_01: Humans beating automation.
• 3:12
  SPEAKER_01: Noted.
• 3:13
  SPEAKER_00: Humans winning on outbound sales is one story.
• 3:15
  SPEAKER_00: The bigger one today is what Meta just gave away for nothing.
• 3:19
  SPEAKER_01: Free as in actually free or free as in sign our terms first.
• 3:23
  SPEAKER_00: Free as in the weights are public.
• 3:26
  SPEAKER_00: Meta released Mew Spark Open yesterday.
• 3:28
  SPEAKER_00: It's an 816 billion parameter model.
• 3:31
  SPEAKER_01: That's enormous.
• 3:32
  SPEAKER_01: How does anyone even run that?
• 3:33
  SPEAKER_00: Here's the trick.
• 3:34
  SPEAKER_00: It's a mixture of experts' design,
• 3:36
  SPEAKER_00: which means only part of the model switches on for any given piece of text.
• 3:40
  SPEAKER_00: Only 42 billion of those parameters activate per token.
• 3:43
  SPEAKER_01: So the rest just sits there, unused.
• 3:46
  SPEAKER_00: Sits in memory, yes, but doesn't do the work.
• 3:49
  SPEAKER_00: The model has 256 specialist subnetworks, and it picks 8 of them for each token it processes.
• 3:55
  SPEAKER_01: And that's light enough to run on what?
• 3:57
  SPEAKER_00: A pair of Apple's M5 Ultra chips, the processor in the top-end Mac Studio.
• 4:02
  SPEAKER_00: You still need serious hardware, just not a data center.
• 4:05
  SPEAKER_01: Is most capable US open-weight model Meta's claim, or somebody else checking?
• 4:10
  SPEAKER_00: That's Meta's own description.
• 4:11
  SPEAKER_00: Write-ups from Hugging Face and independent research outlets confirm the specs.
• 4:16
  SPEAKER_00: Nobody we found has published a head-to-head benchmark against
• 4:18
  SPEAKER_00: the top open models out of China.
• 4:20
  SPEAKER_01: So it might not actually be the best, just the best with that specific label.
• 4:25
  SPEAKER_00: Exactly, and that distinction matters if you're picking a model off the claim alone.
• 4:29
  SPEAKER_00: What's confirmed is the context window, meaning how much text it can hold in memory at once.
• 4:34
  SPEAKER_00: It's a million tokens, roughly a long novel's worth.
• 4:36
  SPEAKER_01: And it takes images, video, audio too?
• 4:39
  SPEAKER_00: As input, yes.
• 4:40
  SPEAKER_00: It's natively multimodal going in, but it only outputs text.
• 4:44
  SPEAKER_00: Feed it a video, get a written answer back.
• 4:46
  SPEAKER_01: Why build it this way, specifically for coding and agents?
• 4:50
  SPEAKER_00: Meta says it's aimed at long multi-step software tasks,
• 4:53
  SPEAKER_00: the kind where a model has to hold context across many steps without losing
• 4:57
  SPEAKER_00: the thread.
• 4:58
  SPEAKER_01: Is it actually out, or is this another coming soon?
• 5:01
  SPEAKER_00: It's out now.
• 5:02
  SPEAKER_00: You can download the weights from Hugging Face or run it
• 5:04
  SPEAKER_00: through API providers like Together AI and Grok starting today.
• 5:07
  SPEAKER_01: So the real headline isn't the size,
• 5:10
  SPEAKER_01: it's that frontier class AI just got a lot cheaper to self-host.
• 5:13
  SPEAKER_00: That's the shift.
• 5:15
  SPEAKER_00: If you've got the hardware,
• 5:16
  SPEAKER_00: there's no per token bill at all,
• 5:17
  SPEAKER_00: which changes the math for anyone currently paying an API per request.
• 5:21
  SPEAKER_01: Until you remember hardware isn't free either.
• 5:24
  SPEAKER_00: Which is exactly the problem underneath today's other AI story.
• 5:28
  SPEAKER_00: Giving the model away doesn't make the chips underneath it free,
• 5:31
  SPEAKER_00: and that's where the price war starts.
• 5:33
  SPEAKER_01: Okay, make the case.
• 5:35
  SPEAKER_01: Why does running AI cost so differently depending on who you ask?
• 5:39
  SPEAKER_00: A new cost analysis,
• 5:40
  SPEAKER_00: using the Artificial Analysis Intelligence Index,
• 5:43
  SPEAKER_00: looked at five frontier models that score within five points of each other
• 5:47
  SPEAKER_00: on benchmarks.
• 5:47
  SPEAKER_01: So basically tied on ability.
• 5:49
  SPEAKER_00: Basically tied.
• 5:50
  SPEAKER_00: But the cost to run one benchmark task ranges from just under $2 to over $7.
• 5:55
  SPEAKER_00: That's a gap of nearly four times.
• 5:57
  SPEAKER_01: Same intelligence, four times the bill.
• 6:00
  SPEAKER_01: Who's cheapest?
• 6:01
  SPEAKER_00: Google's new model, Argonne, prices at $1.99 a task.
• 6:05
  SPEAKER_00: The most expensive in the comparison was Anthropik's Claude Fable 5.1.
• 6:09
  SPEAKER_01: What explains a gap that big if they're equally smart?
• 6:13
  SPEAKER_00: Hardware.
• 6:14
  SPEAKER_00: Google has built its own AI chips alongside its models for 15 years.
• 6:18
  SPEAKER_00: That chip project started back in 2013.
• 6:20
  SPEAKER_01: What kind of chip are we talking about?
• 6:22
  SPEAKER_00: A tensor processing unit, a chip Google designed specifically to run AI workloads.
• 6:27
  SPEAKER_00: Most of its rivals just rent general-purpose graphics chips from NVIDIA instead.
• 6:31
  SPEAKER_01: Does that actually translate into real savings, or is this a just-so story?
• 6:35
  SPEAKER_00: Research on Google's chips found them running up to 70% faster than NVIDIA's equivalent.
• 6:40
  SPEAKER_00: They also use close to half the power.
• 6:42
  SPEAKER_01: That's a real efficiency gap, not just scale.
• 6:45
  SPEAKER_00: Together, that can cut serving costs by as much as 44% at the same profit margin.
• 6:49
  SPEAKER_01: That's a structural advantage, not a marketing number.
• 6:52
  SPEAKER_00: It is, with one catch.
• 6:54
  SPEAKER_00: Google's current price is introductory, and it only runs through the end of November.
• 6:58
  SPEAKER_01: So the $2 price might not last?
• 7:00
  SPEAKER_00: Standard pricing kicks in December 1st.
• 7:03
  SPEAKER_00: At that point, Google lines up with the same rate Anthropic and Meta already charged.
• 7:07
  SPEAKER_00: Some of today's gap is a discount, not just better chips.
• 7:10
  SPEAKER_01: So nobody actually knows yet how much of that four times gap is permanent.
• 7:14
  SPEAKER_00: Not until December 1st, no.
• 7:16
  SPEAKER_00: What is actionable right now is simpler.
• 7:19
  SPEAKER_00: Test your own workload across these models at
• 7:21
  SPEAKER_00: the settings you'd actually use instead of trusting the benchmark ranking alone.
• 7:25
  SPEAKER_01: Because the benchmark says they're tied, but your bill won't be.
• 7:28
  SPEAKER_00: Exactly, and that's the one thing worth doing out of today's news.
• 7:31
  SPEAKER_00: The ranking tells you they're equally capable.
• 7:34
  SPEAKER_00: It tells you nothing about what your specific prompts will actually cost.
• 7:37
  SPEAKER_01: Which matters a lot more once you're paying per token instead of owning
• 7:41
  SPEAKER_01: the hardware, like we just talked about with Meta's model.
• 7:44
  SPEAKER_00: Right, and it's the same economics underneath both stories.
• 7:46
  SPEAKER_00: Somebody pays for the chips either way, whether you self-host a free model or rent an API.
• 7:52
  SPEAKER_00: That's exactly what's rattling investors in the next story.
• 7:54
  SPEAKER_01: Who's actually nervous then?
• 7:56
  SPEAKER_00: Bond investors in CoreWeave, a NeoCloud company.
• 7:59
  SPEAKER_00: That just means a cloud provider that rents out GPU computing power specifically
• 8:03
  SPEAKER_00: for AI workloads, rather than general purpose cloud like AWS.
• 8:07
  SPEAKER_01: And they're worried because?
• 8:08
  SPEAKER_00: CoreWeave has taken on more than $51 billion in debt and obligations
• 8:12
  SPEAKER_00: to build out AI data centers.
• 8:15
  SPEAKER_00: The market just started pricing in real default risk.
• 8:17
  SPEAKER_01: How do you even measure that?
• 8:19
  SPEAKER_00: Is 1100 actually bad or is that normal churn?
• 8:20
  SPEAKER_00: That spread is up around a third in just the past week.
• 8:38
  SPEAKER_00: It's nearly three times what it was a year ago.
• 8:40
  SPEAKER_01: What's the market actually pricing?
• 8:42
  SPEAKER_01: That core weave goes bust or something smaller?
• 8:44
  SPEAKER_00: Specifically,
• 8:45
  SPEAKER_00: that core weave might not generate cash fast enough to service that debt
• 8:49
  SPEAKER_00: on schedule, not that nobody wants its GPUs.
• 8:52
  SPEAKER_01: So is demand actually fine?
• 8:53
  SPEAKER_00: According to analysts covering it, yes.
• 8:56
  SPEAKER_00: CoreWeave already has contracts locked in with major hyperscalers.
• 8:59
  SPEAKER_00: The open question is purely about timing,
• 9:02
  SPEAKER_00: whether the cash from those contracts lands before the bond payments are due.
• 9:05
  SPEAKER_01: That's a pretty narrow worry to spike a credit default swap this hard.
• 9:09
  SPEAKER_00: It's narrow, but it's structural.
• 9:11
  SPEAKER_00: Yields on one of CoreWeave's bonds, due in 2031, have already climbed to nearly 16%.
• 9:16
  SPEAKER_00: That's a steep price to keep borrowing at.
• 9:18
  SPEAKER_01: Is CoreWeave the only company financed this way?
• 9:21
  SPEAKER_00: No, and that's the real significance here.
• 9:23
  SPEAKER_00: A lot of the AI build-out,
• 9:25
  SPEAKER_00: industry-wide,
• 9:26
  SPEAKER_00: runs on debt financing against contracts that assume demand keeps compounding
• 9:30
  SPEAKER_00: on schedule.
• 9:30
  SPEAKER_01: So if CoreWeave wobbles, it's a signal for the whole sector, not just one balance sheet.
• 9:35
  SPEAKER_00: Exactly, that's the reading several analysts are giving it.
• 9:38
  SPEAKER_00: It doesn't mean the AI boom is fake, demand for compute is real.
• 9:42
  SPEAKER_00: It means the financing underneath it is more fragile than the growth story suggests.
• 9:46
  SPEAKER_01: Which ties right back to everything we just said about chips and pricing.
• 9:49
  SPEAKER_00: It does.
• 9:50
  SPEAKER_00: Chip open models and lower per task prices are great for anyone buying AI.
• 9:55
  SPEAKER_00: They're exactly the pressure that makes it harder for the companies financing
• 9:58
  SPEAKER_00: the hardware underneath to make their own numbers work.
• 10:00
  SPEAKER_01: So where does that leave today overall?
• 10:03
  SPEAKER_00: Meta can give away an 816 billion parameter model for free.
• 10:07
  SPEAKER_00: Google can undercut its rivals by almost four times on price.
• 10:11
  SPEAKER_00: And underneath both of those moves,
• 10:12
  SPEAKER_00: the company footing a chunk of the chip bill is starting
• 10:15
  SPEAKER_00: to scare its own bondholders.
• 10:16
  SPEAKER_01: I'd bet that once Google's introductory pricing expires on December 1st,
• 10:21
  SPEAKER_01: at least one other lab cuts its own price to match.
• 10:24
  SPEAKER_01: Not because its costs dropped,
• 10:26
  SPEAKER_01: but because nobody wants to be the expensive option in a market this public
• 10:30
  SPEAKER_01: about pricing.
• 10:31
  SPEAKER_00: December 1st, mark it, that's Friday's news.
• 10:35
  [music]